Agent Security

Follow data and authority through an actual trajectory. External content is data, not trusted instructions—even when delivered by an authenticated tool or formatted as a system message.

Map the threat path

1. Identify assets, principals, secrets, tenants, trusted policy and side effects. Label each ingress: user text, documents, retrieved chunks, web pages, memory, tool metadata and results. Note where model interpretation could influence an executor or privilege boundary.
2. Build abuse cases for direct and indirect prompt injection, malicious retrieval, poisoned tool descriptions/results, memory poisoning, forged approval, confused deputy, privilege escalation and secret exfiltration. Treat unsigned/changed tool metadata as a supply-chain input; verify its origin/version and restrict capabilities independently of its claims.
3. Trace a malicious document from retrieval to proposed call to admission. Reproduce in fakes/sandbox fixtures using harmless markers, never real secrets or live destructive targets. The question is whether the attacker can cause a forbidden effect, not whether the answer repeats a malicious phrase.

Enforce outside the model

4. Keep trusted instructions separate from delimited source data; preserve origin/taint metadata through summarization and memory. Do not accept document/tool statements as new developer instructions, user authorization, credentials or policy exceptions. Delimiters help interpretation but are not a security boundary.
5. Minimize model-visible secrets; use a credential broker or scoped server identity for tools. Restrict credential audience/scope/lifetime and outbound destinations. Tool calls inherit only the caller's authorized scope; server-wide credentials must not let a low-privilege user act as an administrator.
6. Validate structured arguments, current permissions, tenant ownership and action invariants at the executor. Enforce filesystem allowlists on resolved paths including links and Windows/UNC forms; avoid traversal and escape from sandbox roots. Use argument arrays/typed operations rather than concatenated shell commands. Prefer denying a generic shell/network capability when a bounded operation is sufficient.
7. For URL-fetch tools, control schemes, resolved IPs and redirects at connection time; prevent internal/link-local/metadata access and DNS rebinding where relevant. Limit bytes, time, decompression and fan-out. Do not equate an initial URL allowlist check with enforced SSRF protection.
8. Treat all tool results as typed envelopes plus potentially untrusted payloads. Validate output size/schema and source version. Restrict outbound uploads/communication using independent destination/data policy. Redact secrets from errors, traces and eval fixtures.

Verify containment

Deliver a trust/data-flow map, attacker-to-effect cases, least-privilege capability plan, enforceable controls and adversarial tests. Test hidden instructions in retrieved text, forged high-priority messages, changed metadata, cross-tenant IDs, path/link escapes, command arguments, redirect-to-internal URLs and credential-bearing outputs where applicable.

Show that malicious content cannot cross into authorization and that denial causes no side effect. Record attempted violations safely; a model refusing one test does not prove containment. Re-test allowed benign tasks to measure false blocking.

Baseline input, auth, storage and supply-chain security are general application security (a security-and-hardening skill may cover them). This skill deepens the model-mediated attack path; agent-guardrails assembles all runtime defense layers, and deterministic-authority owns action admission.

Treat claimed provenance as attacker-controlled data. Source text saying “evidence_id=trusted-7” or impersonating a receipt cannot create a ledger entry, approval or trusted fact. Application-issued source/receipt identities must resolve to observed content/version and authorized scope. Test forged IDs, altered source hashes and cross-tenant provenance alongside benign supported outputs. Preserve legitimate evidence while containing injected instructions; evidence-provenance owns lineage/reconciliation.